Newsletter

COSMOS welcomes Vida Torgbe as a new Graduate Assistant, where she contributes to research on social media behaviors and dissemination efforts. She is currently pursuing an MS in Information Science at UA Little Rock and brings a multidisciplinary background in social sciences and data analysis. At COSMOS, she works with Prof. Nitin Agarwal to strengthen her skills in social media analytics, data science, and interdisciplinary research while contributing to projects that address complex societal challenges.

Vida shared with us her journey to COSMOS, her goals, and aspirations beyond!

What role do you play at COSMOS?

At COSMOS, I work as a Graduate Assistant, helping Prof. Agarwal with projects focused on understanding social media behaviors. My work involves helping to conduct experiments to evaluate hypotheses. Additionally, I contribute to the center’s research dissemination efforts.

Please share a bit about your professional background and experience.

I have a multidisciplinary background in social sciences and data science, with experience in program evaluation, data analysis, and monitoring and evaluation. I hold a Master of Public Service and am currently pursuing a Master of Science in Information Science at UA Little Rock. My professional experience includes policy research with the Arkansas Department of Education, data analysis with nonprofit organizations, and supporting UNICEF and World Bank social protection initiatives through data-driven monitoring and evaluation. 

What attracted you to join the COSMOS Research Center? What aspects of COSMOS’s vision, mission, and culture stood out to you and why?

What drew me to COSMOS was Prof. Agarwal’s focus on using data science and social media analytics to understand real-world behaviors. Prof. Agarwal’s vision of leading high-impact, multidisciplinary research perfectly aligns with my desire to apply my data science training to meaningful societal issues. It is inspiring to see in the news that Prof. Agarwal and COSMOS are widely recognized as global leaders in social media research, earning international awards for tracking algorithmic bias, combating misinformation, and uncovering narrative manipulation on platforms like YouTube. The culture at COSMOS also stood out to me because it encourages team members from different backgrounds to collaborate and solve these complex problems together.

How do you anticipate your role at COSMOS helping your growth on both a personal and professional level? Are there any specific skills or experiences you’re looking to gain?

At COSMOS, I hope to strengthen my skills in social media analytics, data preprocessing, Python, data visualization, and UI development, as these are core requirements for all COSMOS projects led by Prof. Agarwal. I’m also eager to gain a deeper understanding of tools such as BlogTracker and vTracker and how their data pipelines and interfaces work together. Working under Prof. Agarwal’s guidance will also give me valuable exposure to interdisciplinary research and help me grow as both a researcher and a technical professional.

From your experience, what tips, insights, or advice would you share with someone starting a new role at COSMOS?

My biggest piece of advice would be to fully embrace the collaborative nature of the lab and never hesitate to ask questions. Because COSMOS brings together people from so many different academic backgrounds, there is a wealth of knowledge all around you. A great way to tap into that is to pay attention during the daily huddles when cosmographers share updates on the projects and tasks they are working on. 

If you could share a meal with any historical figure or fictional character, who would it be, and what would you want to talk about and learn from them?

I would choose to share a meal with Maya Angelou. Having worked closely with the Celebrate Maya Project in Little Rock, I have developed a deep appreciation for her legacy, resilience, and profound ability to capture the human experience. During our conversation, I would love to connect my background in public service and a strong desire for social media analysis with her unique perspective on community engagement. I would want to talk to her about how we can best use modern data tools and technology to uncover societal needs, tell authentic stories, and shape public policies that protect vulnerable populations.

In this month’s research spotlight, COSMOS highlights three studies presented at the 14th International Conference on Complex Networks and their Applications in New York, USA, exploring the underlying mechanics of digital influence, grievance clustering, and narrative transmission. When major geopolitical events unfold, social media platforms function as dynamic information ecosystems where personal reactions can swiftly evolve into structured movements. Understanding how everyday concerns translate into connected communication networks, which policy storylines achieve sustained reach, and how visual symbols drive engagement is crucial for deciphering modern collective behavior.

The first study, “The Network Effect of Shared Grievances: Measuring Collective Concern of Tariff Policy,” examines how tariff-related grievances mobilize online communities. Using a multi-stage framework that integrates GPT-4o-mini classification, HDBSCAN clustering, and user-mention networks across six months of data, the authors analyzed the structural divergence between consumer and economic grievances. The results reveal that tangible consumer hardships form broad, overlapping clusters with fluid boundaries that experience sudden, synchronized surges around external triggering events. In contrast, macroeconomic concerns organize into dense, specialized clusters centered on specific trade sectors and policy figures, demonstrating how online grievance networks form the pre-mobilization groundwork for broader collective action.

The second study, “Modeling the Propagation Dynamics of Visual Elements with Epidemiological Frameworks,” introduces a novel computational approach to track how visual symbols spread across video platforms during information manipulation in Taiwan. Utilizing the PRISM color-shift model to extract key video frames alongside vision-language models for symbol classification, the authors evaluated five epidemiological models to trace symbol dissemination across nearly 2,000 YouTube videos. The SEIZ framework, which explicitly accounts for an undecided or skeptical audience compartment, achieved the highest fidelity by dropping modeling error to 0.45%. Crucially, the findings show that transmission effectiveness does not depend on symbol saturation, but rather on strategic pairing, with political symbols and dual-symbol combinations achieving near-optimal viral spread.

The third study, “How Tariff War Discourse Spreads on Social Media? A Study of Narrative Outbreak,” bridges qualitative narrative analysis with mathematical epidemiology to track competing storylines during international trade disputes. By clustering multilingual Twitter discourse and applying GPT-4o with structured prompts, the authors extracted five core narratives and modeled their transmission rates using bounded parameter optimization algorithms, including Nelder-Mead and L-BFGS-B. The analysis found that analytically grounded narratives, specifically those focused on reciprocal tariffs, strategic retaliation, and institutional trade responses, consistently demonstrated the highest transmissibility and lowest error. Meanwhile, emotionally charged and satirical narratives exhibited volatile decay, proving that policy-rich, expert-amplified content sustains longer digital lifespans.

Together, these studies advance COSMOS’s mission to build robust, interpretable approaches for analyzing complex digital ecosystems. For science, they bridge mathematical diffusion models with qualitative narrative extraction and AI-enhanced computer vision, offering scalable methodologies to audit information flows. For society, they provide platform owners, researchers, and policymakers with actionable diagnostic tools to identify emerging collective narratives, evaluate visual campaigns, and anticipate the trajectory of contentious public discourse.

COSMOS Research Center is pleased to announce a new publication in Springer’s Journal of Social Network Analysis and Mining titled “Identifying Cohesive Narrative Formations in Cross-Platform Trade-War Discourse Using Weighted Contextual Focal Structures Analysis.”

Geopolitical discourse across social media platforms rarely unfolds in isolated silos. Instead, complex storylines emerge through the dynamic interplay of user interactions, semantic alignment, and cross-platform mechanics. Traditional Contextual Focal Structure Analysis (CFSA) models have helped identify context-specific user groups, but they treat all connections equally and analyze apps separately. This makes it difficult to capture subtle semantic nuances, stance homogeneity, or varying interaction intensities across platforms. The study addresses these issues by introducing Weighted Contextual Focal Structures Analysis (W-CFSA), a novel framework that evaluates multiple platforms at once. By combining custom platform weights, interaction frequency, and text similarity, W-CFSA analyzes the entire network layout to identify highly connected groups that share the same core message.

Using a dataset of 73,266 posts collected from YouTube, TikTok, X (formerly Twitter), and Instagram during the 2025 U.S.-China trade policy discussions, the researchers systematically benchmarked W-CFSA against size-matched eigenvector-centrality baseline groups. The findings demonstrated that W-CFSA substantially outperformed standard baseline methods across all structural metrics. The identified focal sets maintained consistently higher internal weighted density and clustering coefficients while exerting a far greater impact on global network transitivity when suspended. Qualitative analysis further revealed that these focal sets correspond to distinct narrative architectures like specific causal interpretations of policy volatility or strategic leverage, rather than simple stance or sentiment categories.

This research advances computational social science by offering trust, safety, and strategic communication teams a methodological framework to trace cross-platform narrative propagation at its core. By bridging network topology with semantic context, COSMOS continues to pave the way for a deeper understanding of digital discourse and information ecosystems.

Click here to read the full article.

We are thrilled to announce that three new COSMOS studies have been accepted and presented at the 32nd Americas Conference on Information Systems (AMCIS 2026) in Reno, Nevada.

Organized by the Association for Information Systems, AMCIS brings together leading researchers examining how interface design and algorithms actively shape human judgment. It serves as a vital venue for dissecting the friction between platform mechanics and public understanding, particularly around toxic evasion, recommendation rabbit holes, and the cultural cues that anchor voter sentiment. Presenting our work here reinforces COSMOS’s commitment to building transparent and auditable AI tools that help communities navigate increasingly complex digital spaces.

Presenting at AMCIS directly reflects our work supported by federal grants dedicated to national defense, platform accountability, and socio-technical resilience. Led by Prof. Nitin Agarwal, these projects examine how users experience the internet during volatile moments, connecting the dots between algorithmic architecture and human cognition. Through this research, our team analyzes hundreds of thousands of user interactions across high-stakes arenas, including polarized debates on Reddit, multi-hop recommendation trails on YouTube during international trade disputes, and information manipulation efforts in Taiwan.

The three studies presented at AMCIS showcase COSMOS’s interdisciplinary approach, which fuses social computing and Artificial Intelligence (AI)-based techniques to diagnose and mitigate risks across digital platforms. The first study, led by Prof. Agarwal, investigates toxic algospeak across 13,780 high-evasion Reddit posts out of 721,236 climate discussions, uncovering a dual-mechanism cognitive framework where surface-level euphemisms trigger universal metalinguistic detection without harm discounting, whereas domain-specific jargon misuse requires epistemic sophistication, which enables experts to discount perceived harm by 16.5%. This distinction proves that one-size-fits-all moderation fails and offers practical governance guidance to separate user education from expert-led validation. 

The second study, led by Prof. Agarwal, introduces TrapIntensity, an auditable sociotechnical framework that measures AI algorithmic entrapment in YouTube recommendation graphs by pairing hop-aware random-walk network simulations with theory-driven persuasion cue extraction using large language models. This dual-layer metric reveals whether engagement concentration is driven by structural lock-in or rhetorical persuasion across different contexts, showing structural dominance in trade dispute networks and persuasion dominance in socio-political discourses. 

The third study examines a multimodal dataset of 1,973 YouTube videos and over 342,000 comments from Taiwan’s information environment, applying the PRISM perceptual keyframe framework alongside vision-language models to show that social, cultural, and political symbols systematically drive higher viewer interaction in information campaigns, with cultural symbols eliciting the strongest emotional resonance and driving expressed trust in Taiwanese institutions to nearly 45%. 

Together, these publications reinforce COSMOS’s continued mission to explain, measure, and mitigate complex digital challenges by bridging advanced AI methodologies with human-centered socio-technical systems.

COSMOS Research Center has been honored with the prestigious Best Paper Award at the International Conference on AI-based Media Innovation (AIMEDIA 2026), held July 5 to 9, 2026, in Nice, France! Hosted by the International Academy, Research, and Industry Association (IARIA), AIMEDIA serves as an international forum dedicated to evaluating artificial intelligence across media ecosystems, addressing emerging challenges related to information integrity, algorithmic bias, and ethical media innovation.

The research, titled “Symbolic Communication in the 2025 Tariff Discourse: A Comparative Analysis of X and Weibo”, led by Prof. Nitin Agarwal, examines how international trade conflicts are discussed online across different cultures and digital spaces, focusing specifically on the 2025 trade/tariff wars. By analyzing thousands of posts on X and Weibo, we investigated the drivers of public interest and engagement during global economic disputes. The team used three distinct artificial intelligence systems as digital interpreters to categorize the underlying themes of each post as social, cultural, economic, or political.

The findings reveal two distinct online environments with markedly different communication styles. Discussions on X tend to be straightforward, narrow, and primarily focused on financial figures, with most posts sticking to a single topic. On the other hand, conversations on Weibo are highly energetic and multi-layered, frequently blending multiple themes at once and heavily relying on cultural symbols, metaphors, and emotional and nationalistic expressions.

By applying mathematical models typically used by epidemiologists to track how viruses spread, the study showed that Weibo discussions spread significantly faster and wider than those on X. Across both platforms, the research highlights that technical policy debates only truly capture the public’s imagination when they are framed with cultural and emotional resonance. This demonstrates that to understand the public impact of international trade policies, we must look beyond spreadsheets and pay close attention to the cultural stories people share.

Where did you complete your internship, and what was the primary focus of your projects?

I did my summer internship at Meta, Inc. (parent company of Facebook), where I worked on developing “Early Stage Ranking Models” for Product Centric Ads (Recommendation Systems)

What skills did you develop, what challenges did you face, and what was your most valuable takeaway from the experience?

During my internship, I developed skills in ranking systems, machine learning experimentation, model evaluation, and production-focused AI. The main challenge was learning how to build under real production constraints, such as scale and latency. My biggest takeaway was that strong ML solutions must be accurate, efficient, and practical for real-world deployment. This experience showed me how machine learning research becomes useful when it can work reliably in a real production system.

In what ways did your COSMOS research experience prepare you for your internship?

I work as a graduate assistant at the COSMOS Research Center under Prof. Nitin Agarwal’s supervision on projects that honed my skills in social computing, behavioral modeling, machine learning, AI, large-scale data analysis, and model evaluation. Prof. Agarwal is also my doctoral advisor and dissertation chair. His mentorship enabled me to conduct competitive, team-oriented, and application-driven research. He encourages us to publish at top-tier venues. These skills helped me tremendously during my internship at Meta, a highly competitive and mission-driven environment.

How has your internship experience enhanced your current work and research at COSMOS?

My internship influenced how I think about our research at COSMOS. It helped me see the value of building models that are not only accurate but also efficient, scalable, and useful in real-world applications. I now bring this practical view into the research project led by Prof. Agarwal at the COSMOS Research Center.

How did your internship broaden your perspective on the type of research conducted at COSMOS? 

My internship broadened my view of research conducted at COSMOS. At COSMOS, we study how user connections on social media shape social behavior from outside the platforms. At Meta, I studied these behaviors from within the platform. Specifically, I analyzed how similar user-item and user-content connections can be used to support recommendation systems. This helped me see that the same research ideas can apply to both social analysis and product systems. Further, I recognize the need for an application-oriented approach to research – something that Prof. Agarwal consistently champions.

What advice would you offer to fellow COSMOS researchers preparing for future internships?

Be steadfast. Work hard. Learn as much as you can. Know your research and its value to society. No shortcuts. Listen to Prof. Agarwal!

Share a memorable story or moment from your internship that stood out to you.

One memorable moment was when one of my peer leads went on a refresh, and I had to help set up a war room with senior engineers. It was an eye-opening experience because I had to respond quickly, organize the work, communicate clearly, and solve technical problems under pressure. The moment tested my engineering, leadership, and interpersonal skills.

If you had to describe your internship experience in one word, what would it be? 

Incredible!

In this month’s research spotlight, COSMOS highlights three studies published at the 14th International Conference on Complex Networks and their Applications (Complex Networks), held in New York, USA, that investigate how recommendation systems shape user attention, content visibility, and online behavior. As users move from one recommended video to another, digital platforms can quietly nudge them toward tightly connected clusters of content. These clusters, often described as content traps, may narrow exposure, reinforce specific viewpoints, and steer attention towards (or away from) certain narratives or topics.

The first study, “How Far is Too Far? Modeling User Attraction Pathways in Recommendation Networks via Random Walk Variants,” examines how easily users can encounter structurally influential groups within a YouTube recommendation network. Using hop-aware random walk simulations, the study models how users may move from different distances in the network and compares neutral exploration with popularity-driven navigation. The study shows that certain focal structures are more reachable than other network groupings, offering insight into how recommendation pathways can make some content clusters more visible than others.

The second study, “TrapIntensity: Quantifying Structural Entrapment via Hop-Aware Attraction and Retention,” builds on this idea by asking not only whether users can reach a content cluster, but also how strongly that cluster can hold attention. The framework combines attraction and retention into a unified trap intensity score, helping identify network regions that are both easy to enter and difficult to leave. This offers a more interpretable way to study content traps, echo chambers, and filter bubbles in recommendation systems.

The third study, “Persuasive Pathways into Content Traps: The Role of Persuasive Features in Structuring Algorithmic Content Cycles,” looks beyond network structure to examine the content itself. The research investigates how persuasive features in YouTube transcripts interact with topical uniformity and engagement. The study finds that highly homogeneous content groups tend to contain stronger persuasive signals and higher engagement, suggesting that content traps are not only structural but also rhetorical. In other words, users may remain in these cycles not just because of how recommendations are connected, but because the content itself is persuasive and reinforcing.

Together, these studies tell a broader story about algorithmic influence. AI-based recommendation systems do more than suggest content; they shape pathways of influence and attention. By combining network science, random walk modeling, persuasion theory, and engagement analysis, our research advances new ways to understand how content traps form, why they persist, and how they can be studied more transparently. Collectively, it reflects COSMOS’s mission to develop robust, interpretable, and socially meaningful approaches for analyzing digital ecosystems. For science, it contributes new methods for auditing recommendation networks and modeling user exposure. For society, it supports a deeper understanding of how online platforms and their AI-based algorithms can influence information diversity, user agency, and the dynamics of digital behavior.

COSMOS Research Center is pleased to announce a new publication in Springer’s Journal of Social Network Analysis and Mining titled “Large-Scale Toxicity Intervention in Social Networks: Evaluating Integer Programming-Optimized Focal Toxic Structures”.

Online harassment and toxic campaigns rarely happen in isolation. Instead, they are driven by focal toxic structures (densely connected groups of social media accounts) that work together to amplify harmful content and evade standard filters. Traditional content moderation systems focus on removing individual posts or banning single accounts, but these coordinated groups simply create new profiles or use coded language to keep campaigns alive. Platform moderation teams face strict resource limits and cannot review every flagged user, making it critical to know exactly which groups to prioritize for removal. To solve this, the study introduces an advanced optimization model combining Weighted Focal Structure Analysis with Integer Programming optimization. Instead of evaluating toxic groups one by one, this mathematical approach looks at the entire network at once to select the most impactful combination of groups to remove. It maximizes network disruption while respecting real-world constraints, such as moderation budget limits, overlap prevention, and minimum impact thresholds.

Using a large-scale dataset of 324,769 Telegram users active during the Russia-Ukraine conflict, the research systematically compared standard sequential selection against the new optimized approach and a hybrid model. The findings revealed that the mathematically optimized approach drastically outperformed standard baseline methods across every metric. It achieved up to a 149% increase in network fragmentation, split up the main toxic highways twice as effectively, and achieved nearly double the reduction in overall harmful content while removing 17% fewer total users.

This research advances computational social science by giving trust and safety teams a scalable, data-driven framework to dismantle harmful campaigns at their root. By shifting the focus from reactive post removals to strategic network interventions, COSMOS continues to pave the way for safer, more resilient digital ecosystems.

Click here to read the full article.

COSMOS Research Center is proud to highlight its recent presentation of three milestone studies at the Cyber Social Threats (CySoc), held at the 20th International AAAI Conference on Web and Social Media (ICWSM 2026) in Los Angeles, USA.

Organized by the Association for the Advancement of Artificial Intelligence (AAAI), ICWSM is recognized internationally as a flagship venue for computational social science, bridging advanced data science with human behavior analysis. Its specialized CySoc workshop provides an essential platform for addressing the dark side of digital platforms, focusing on cyber social threats, coordinated online manipulation, and digital behavior during acute political crises. Participating in this highly competitive venue highlights COSMOS’s ongoing commitment to advancing data-driven tools that safeguard public discourse and foster digital resilience.

Publishing at AAAI ICWSM CySoc directly advances COSMOS’s mission to analyze online influence, cognitive security, and information stability, supported in part by major federal grants dedicated to understanding dynamics across strategic regions, including the Indo-Pacific. Through these projects, COSMOS researchers examine real-world social movements, cross-platform behaviors, and digital discourse across key international contexts, including Nepal, Taiwan, and broader global trade ecosystems.

The three studies presented at CySoc showcase COSMOS’s multi-pronged approach to modeling and mitigating digital threats. The first study extends classical epidemiological frameworks by introducing an immediate relapse mechanism into the SEIRS model, successfully capturing how toxic behavior repeatedly flares up in online communities. The second paper analyzes multilingual discourse from the 2025 youth-led protests in Nepal across five temporal crisis phases, revealing a critical finding for platform moderation: while switching languages within reply chains often reduces general personal insults, it selectively elevates physical threat language, uncovering a covert threat vector that standard filters overlook. The third study applies Weighted Focal Structure Analysis (WFSA) to YouTube recommendation graphs during major events like the 2024 Taiwan presidential election and global tariff debates, demonstrating how algorithmic ranking and navigation depth concentrate structural power within specific content pathways.

Together, these studies highlight COSMOS’s international leadership in socio-cognitive security and ethical AI. By combining advanced mathematical modeling, natural language processing, and network science, COSMOS continues to deliver scalable, real-world insights to counter emerging cyber social threats across the globe.

In a recent Q&A feature published by Apprenticely, COSMOS Research Center director Prof. Nitin Agarwal reflected on a milestone judicial verdict holding tech giants Meta and YouTube liable for negligent, addictive platform design. The landmark ruling brings sweeping scrutiny to how major social media platforms operate, shifting public and legal focus toward the engineered mechanisms that drive modern digital environments.

Prof. Agarwal explained that social media algorithms are fundamentally optimized for attention maximization and engagement rather than user well-being. Powered by AI recommender systems, platforms leverage psychological triggers such as variable rewards, social validation, and the fear of missing out (FOMO). Features like infinite scroll and continuous autoplay intentionally strip away natural stopping cues, creating repetitive feedback loops that reinforce compulsive behavior. Because time-on-platform directly translates into advertising revenue, user retention is not an unintended side effect, but rather an explicitly engineered outcome.

Beyond individual psychological impacts, COSMOS’s research demonstrates how these platform vulnerabilities scale socially and politically. Supported by over $30 million in project funding from federal agencies, including the U.S. Department of War and the National Science Foundation, COSMOS studies how algorithms shape collective human behavior during crises and socio-political events. Adversarial actors routinely exploit these engagement-first models by deploying bots, coordinated posting, and outrage-driven content to game recommendation systems. By weaponizing human psychology and platform design, malicious actors effectively turn content curation algorithms into vectors for large-scale cognitive influence and disinformation.

Comparing these recent legal developments to the historical scrutiny faced by Big Tobacco, Prof. Agarwal outlined a potential path forward for tech regulation. As internal corporate knowledge and design intents come under increased judicial review, platforms may face mandatory duty-of-care standards, age protections for minors, and structural requirements for algorithmic transparency. Highlighting Arkansas Governor Sarah Huckabee Sanders’s initiative to tackle youth exposure and online harms, Prof. Agarwal noted growing momentum among policymakers. Serving as a member of the Governor’s AI Taskforce, he suggested that future AI systems could be reoriented away from pure attention harvesting and toward healthy engagement, fostering a more accountable and resilient digital ecosystem. 

Click here to read the full Q&A feature on Apprenticely.